The HTTP API

EveryAPI speaks three request dialects on one host: OpenAI (Chat Completions and Responses), Anthropic Messages, and Google Gemini. Pick whichever your SDK already emits — the gateway converts to the upstream provider's format and converts the response back.

Base URL and auth

Base URL   https://api.everyapi.ai          global
           https://api-cn.everyapi.ai       China-accelerated
Auth       Authorization: Bearer sk-everyapi-…

The Anthropic and Gemini surfaces also accept their native auth headers — x-api-key plus anthropic-version for Anthropic, x-goog-api-key or a ?key= query parameter for Gemini — carrying the same sk-everyapi-… value. That is what lets an unmodified vendor SDK point at the gateway.

OpenAI-compatible endpoints

POST   /v1/chat/completions      Chat Completions
POST   /v1/completions           Legacy completions
POST   /v1/responses             Responses API
POST   /v1/responses/compact     Responses compaction
DELETE /v1/responses/{id}        Delete a stored response
POST   /v1/embeddings            Embeddings
POST   /v1/rerank                Rerank
POST   /v1/moderations           Moderation
POST   /v1/images/generations    Image generation
POST   /v1/images/edits          Image editing
POST   /v1/audio/speech          Text to speech
POST   /v1/audio/transcriptions  Speech to text
POST   /v1/audio/translations    Speech translation
GET    /v1/realtime              Realtime (WebSocket)
GET    /v1/models                Model catalog
GET    /v1/models/{model}        One model

Anthropic-compatible endpoints

POST   /v1/messages               Messages
POST   /v1/messages/count_tokens  Token counting (relayed, not billed)

/v1/models returns the Anthropic-shaped catalog when the request carries both x-api-key and anthropic-version.

Gemini-compatible endpoints

POST   /v1beta/models/{model}:{action}   Native generate / stream / embed
GET    /v1beta/models                    Gemini-shaped catalog
GET    /v1beta/openai/models             OpenAI-shaped catalog, Gemini path

Async media: submit then poll

Video and music are long-running, so they are submit-and-poll task channels rather than a single blocking call.

POST   /v1/video/generations            Submit (OpenAI-compatible)
GET    /v1/video/generations/{task_id}  Poll
POST   /v1/videos                       Submit (Sora-style)
GET    /v1/videos/{task_id}             Poll
POST   /v1/videos/{video_id}/remix      Remix an existing result
GET    /v1/videos/{task_id}/content     Fetch the finished media
DELETE /v1/videos/{task_id}             Cancel

Native provider surfaces exist alongside the compatible one, for clients that already speak them:

POST /kling/v1/videos/text2video            Kling, native shape
POST /kling/v1/videos/image2video
GET  /kling/v1/videos/text2video/{task_id}
GET  /kling/v1/videos/image2video/{task_id}
POST /jimeng/                Jimeng, official Action-parameter shape
POST /suno/submit/{action}   Suno music and lyrics
POST /suno/fetch             Suno batch poll
GET  /suno/fetch/{id}
POST /mj/submit/{action}     Midjourney (imagine, blend, describe, …)
GET  /mj/task/{id}/fetch     Midjourney poll

Supported async video providers include Sora, Kling, Hailuo, Jimeng, Vidu, and Doubao; Suno covers music.

Streaming

Set "stream": true and read standard SSE: data: <json> frames terminated by data: [DONE]. The Anthropic and Gemini surfaces stream in their own native event shapes. Realtime uses a WebSocket at /v1/realtime.

Endpoints that return 501

These exist in the OpenAI spec but are deliberately not proxied. They answer HTTP 501 with code=endpoint_not_supported:

POST   /v1/images/variations
GET    /v1/files          POST /v1/files
GET    /v1/files/{id}     DELETE /v1/files/{id}
GET    /v1/files/{id}/content
GET    /v1/fine-tunes     POST /v1/fine-tunes
GET    /v1/fine-tunes/{id}
POST   /v1/fine-tunes/{id}/cancel
GET    /v1/fine-tunes/{id}/events
DELETE /v1/models/{model}

Provider file storage and fine-tune lifecycles are too provider-specific to normalise at a gateway, and model deletion would conflict with the marketplace's channel ownership. For these, call the upstream provider directly with that provider's own credentials.

Error responses and language

Errors come back in the OpenAI error envelope. Backend messages are translated: send Accept-Language and you get English, Simplified Chinese, Traditional Chinese, Japanese, Korean, Spanish, German, or French. The CLI sets this automatically from your language setting.

Rate limits and quota

Per-user, per-model rate limits apply on top of the per-key quota. A key can additionally be restricted to a model list, a routing group, and an IP allowlist — see the tokens topic. Requests that exceed a limit return 429; requests with no quota left fail before reaching an upstream.